A machine learning-based method for defending a voice assistant from being controlled by a silent instruction, comprising the following steps: 1) collecting data of positive and negative samples; 2) performing data sharding on the data of the positive and negative samples; 3) selecting and normalizing sample features; 4) selecting a classifier to be trained and generating a detection model for a malicious voice instruction; and 5) detecting a voice instruction to be detected by using the obtained detection model. For intelligent devices of different types, it is necessary to obtain a normal voice instruction and a malicious voice instruction by means of an intelligent device of a specific type, and using the normal voice instruction and the malicious voice instruction as positive and negative samples to train a specific classifier for the device of the type. Such a customized approach can well solve the problem that detection and defense between devices cannot work.